Fix Kimi Coding tool-call replay by amknight · Pull Request #82550 · openclaw/openclaw

amknight · 2026-05-16T11:22:02Z

Summary

Fixes #82161 by preserving Kimi Coding reasoning replay on OpenAI-compatible follow-up turns and by backfilling the blank reasoning_content placeholder Kimi expects on assistant tool-call replay messages when thinking is enabled.

Reproduction

I reproduced the issue against current main with a source-level probe for { provider: "kimi", api: "openai-completions", id: "kimi-for-coding" }. The failing behavior was that buildOpenAICompletionsParams serialized the assistant tool-call replay without reasoning_content, even though the Kimi/Moonshot-compatible K2.6 route preserved it.

I also ran the same reproduction path in Blacksmith Testbox before the fix. That Blacksmith source probe showed assistantHasReasoningContent: false for kimi-for-coding. The Blacksmith environment did not expose a live Kimi/Moonshot key, so live provider auth was blocked; the manual E2E proof below uses a local Kimi-compatible endpoint that rejects the same missing-field condition.

Root Cause

kimi-for-coding was not in the OpenAI-compatible reasoning replay allowlist used by the Chat Completions payload sanitizer, so the replay sanitizer stripped provider-owned reasoning_content before the next request. The Kimi provider wrapper also did not add the blank reasoning_content placeholder for assistant tool-call replay messages when OpenAI-compatible Kimi thinking was enabled.

Fix

Adds kimi-for-coding to the replay-required model ids in the OpenAI Chat Completions sanitizer and transcript policy.
Expands model-id matching to handle prefixed/suffixed aliases such as hf:moonshotai/kimi-k2-thinking and xiaomi/mimo-v2.6-pro:cloud.
Updates the Kimi Coding stream wrapper to backfill reasoning_content: "" on assistant tool-call replay messages when thinking is enabled, and to strip replay reasoning fields when thinking is disabled.
Adds focused regression coverage for transport serialization, transcript policy, and the Kimi wrapper.
Adds a changelog entry.

Verification

node scripts/run-vitest.mjs src/agents/openai-transport-stream.test.ts src/agents/transcript-policy.test.ts extensions/kimi-coding/stream.test.ts locally after rebasing onto origin/main (400 passed).
git diff --check origin/main...HEAD.
codex review --uncommitted before publishing: no actionable correctness findings.

Real behavior proof

Behavior addressed: Kimi Coding OpenAI-compatible follow-up turns after tool use must include provider-owned assistant replay reasoning_content so Kimi does not reject the next request.

Real environment tested: Blacksmith Testbox tbx_01krr7g58dy8f2qcxd7vqxh01q, GitHub Actions run https://github.com/openclaw/openclaw/actions/runs/25960370706.

Exact steps or command run after this patch: node scripts/crabbox-wrapper.mjs run --provider blacksmith-testbox --blacksmith-workflow .github/workflows/ci-check-testbox.yml --blacksmith-job check --blacksmith-ref main --timing-json --shell -- "... node scripts/run-vitest.mjs src/agents/openai-transport-stream.test.ts src/agents/transcript-policy.test.ts extensions/kimi-coding/stream.test.ts; ... manual Kimi-compatible loopback E2E ...".

Evidence after fix: Blacksmith focused tests passed (5 passed, 396 passed at that pre-rebase base), then the manual E2E fake provider captured /v1/chat/completions with thinking: { "type": "enabled" } and assistant replay reasoning_content: "".

Observed result after fix: the fake Kimi-compatible endpoint returned KIMI_REPLAY_OK; the same harness rejects the request if assistant tool-call replay is missing reasoning_content.

What was not tested: a live Kimi API request, because the Blacksmith environment did not provide KIMI_API_KEY, KIMICODE_API_KEY, or MOONSHOT_API_KEY.

clawsweeper · 2026-05-16T11:25:51Z

Codex review: passed.

Summary
The PR preserves Kimi Coding reasoning_content replay for OpenAI-compatible tool-call follow-up turns, extends replay model-id matching, adds Kimi wrapper/tests, and updates the changelog.

Reproducibility: yes. at source level: current main drops or fails to synthesize reasoning_content for kimi-for-coding replay, and the linked report includes a concrete Kimi 400 after tool-call history. I did not run a live Kimi request in this read-only review.

Real behavior proof
Not applicable: The external contributor proof gate does not apply to this MEMBER/maintainer-labeled PR; the body still includes Testbox and loopback endpoint proof, but no live Kimi API run.

Next step before merge
No repair lane is needed; the PR is already opted into automerge and has no actionable review findings, so checks and mergeability should gate it.

Security
Cleared: The diff changes provider payload serialization, tests, and changelog text only; it adds no dependencies, workflow execution, secrets handling, or package-resolution behavior.

Review details

Best possible solution:

Land the narrow replay serialization fix once exact-head checks and merge gates pass, then split any remaining Kimi tool-call argument assembly failure into its own focused follow-up.

Do we have a high-confidence way to reproduce the issue?

Yes, at source level: current main drops or fails to synthesize reasoning_content for kimi-for-coding replay, and the linked report includes a concrete Kimi 400 after tool-call history. I did not run a live Kimi request in this read-only review.

Is this the best way to solve the issue?

Yes for the replay serialization bug: the PR extends provider-specific replay policy and the Kimi wrapper without weakening stock OpenAI sanitization. The later argument-assembly report should be a separate follow-up if it still reproduces.

What I checked:

Protected PR state: Live PR API shows author_association MEMBER, labels include maintainer and clawsweeper:automerge, head 9a4605e, and the PR is open/mergeable. (9a4605ee3843)
Current main transport gap: Current main preserves reasoning_content only for compat flags or listed model ids; the allowlist lacks kimi-for-coding and the normalizer splits at the first colon before slash handling. (src/agents/openai-transport-stream.ts:2552, 9b560b8a41f4)
Current main transcript-policy gap: Current main's strict OpenAI-compatible transcript policy lacks kimi-for-coding in the replay-required model-id set, so it can drop historical reasoning for that route. (src/agents/transcript-policy.ts:160, 9b560b8a41f4)
Current main Kimi wrapper gap: The Kimi stream wrapper maps thinking into the provider payload and strips generic reasoning knobs, but it does not backfill assistant tool-call reasoning_content for OpenAI-compatible replay. (extensions/kimi-coding/stream.ts:321, 9b560b8a41f4)
PR diff matches implicated paths: The patch adds kimi-for-coding to both replay allowlists, broadens prefixed/suffixed model-id matching, and adds Kimi wrapper logic plus focused regression tests for enabled and disabled thinking payloads. (extensions/kimi-coding/stream.ts:75, 9a4605ee3843)
Related issue context: The linked issue gives a concrete 2026.5.12 Kimi 400 for missing reasoning_content and later adds a separate possible tool-call argument assembly failure, so this PR addresses the replay half but may not close every reported symptom.

Likely related people:

steipete: Recent commits preserved Kimi/MiMo reasoning replay and standardized the Kimi Coding model id near the affected transport and provider catalog surfaces. (role: recent provider replay and Kimi provider contributor; confidence: high; commits: 99dfb8291f1c, 2d8339529b58, 53cd3daed3cb; files: src/agents/openai-transport-stream.ts, src/agents/transcript-policy.ts, extensions/kimi-coding/provider-catalog.ts)
jimdawdy-hub: Recent transport work changed requiresReasoningContentOnAssistantMessages handling, the replay mechanism this PR extends. (role: transport compat contributor; confidence: high; commits: d00cd314d17b, d4c83edba89f; files: src/agents/openai-transport-stream.ts, src/agents/openai-completions-compat.ts)
Dan O'Brien: Prior Kimi thinking stream work is directly adjacent to the wrapper changed by this PR. (role: Kimi thinking stream contributor; confidence: medium; commits: 8f173563926b; files: extensions/kimi-coding/stream.ts)
frankekn: Recent Kimi thinking-default work is adjacent to enabled/disabled thinking payload behavior in the same plugin area. (role: Kimi thinking behavior contributor; confidence: medium; commits: 4ca5f51430d7; files: extensions/kimi-coding/stream.ts)

Remaining risk / open question:

The PR body has Testbox and loopback payload proof, but no live Kimi API request because the environment lacked Kimi/Moonshot credentials.
The linked issue includes a later possible tool-call argument assembly failure that this replay-focused patch does not prove or fix.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 5a14b1c5c5d7.

amknight · 2026-05-16T11:29:58Z

/clawsweeper automerge

clawsweeper · 2026-05-16T11:31:49Z

🦞✅
ClawSweeper merged this PR after the passing review.

Source: clawsweeper[bot]
Feedback: structured ClawSweeper verdict: pass (sha=9a4605ee3843e2fdb9dd8d61b39400e58871923a)
Merge status: merged by ClawSweeper automerge
Merged at: 2026-05-16T11:54:47Z
Merge commit: f8b7008f7c53

What merged:

The PR preserves Kimi Coding reasoning_content replay for OpenAI-compatible tool-call follow-up turns, extends replay model-id matching, adds Kimi wrapper/tests, and updates the changelog.
Reproducibility: yes. at source level: current main drops or fails to synthesize reasoning_content for kimi- ... es a concrete Kimi 400 after tool-call history. I did not run a live Kimi request in this read-only review.

Automerge notes:

No ClawSweeper repair was needed after automerge opt-in.

The automerge loop is complete.

Automerge progress:

2026-05-16 11:29:58 UTC review queued fcc577e85d93 (queued)

2026-05-16 11:31:48 UTC repair queued fcc577e85d93 (autonomous) Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068

2026-05-16 11:34:10 UTC repair started (running) in 1s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 automerge-openclaw-openclaw-82550

2026-05-16 11:34:26 UTC validation plan (passed) in 16s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 pnpm check:changed; pnpm lint; pnpm check:test-types

2026-05-16 11:34:38 UTC Codex write preflight (passed) in 28s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 danger-full-access

2026-05-16 11:43:04 UTC Codex edit 1 7d143a68f263 (complete) in 8m 55s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 exit 0

2026-05-16 11:41:07 UTC review passed fcc577e85d93 (structured ClawSweeper verdict: pass (sha=fcc577e85d935692b7b5fdf391c62cde57533...)

2026-05-16 11:46:39 UTC validation and review 1 9a4605ee3843 (base moved) in 12m 30s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 rebased

2026-05-16 11:46:45 UTC repair completed 9a4605ee3843 (branch updated) in 12m 35s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 initial automerge rebase is delegated to Codex repair

2026-05-16 11:46:44 UTC review queued 9a4605ee3843 (after repair)

2026-05-16 11:54:43 UTC automerge wait 9a4605ee3843 (ready) in 20m 33s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 checks and exact-head review are ready

2026-05-16 11:54:44 UTC repair finished 9a4605ee3843 (pushed) in 20m 34s Run: https://github.com/openclaw/clawsweeper/actions/runs/25960832068 repair_contributor_branch

2026-05-16 11:54:33 UTC review passed 9a4605ee3843 (structured ClawSweeper verdict: pass (sha=9a4605ee3843e2fdb9dd8d61b39400e588719...)

2026-05-16 11:54:49 UTC merged 9a4605ee3843 (merged by ClawSweeper automerge)

Summary: - The PR preserves Kimi Coding reasoning_content replay for OpenAI-compatible tool-call follow-up turns, extends replay model-id matching, adds Kimi wrapper/tests, and updates the changelog. - Reproducibility: yes. at source level: current main drops or fails to synthesize reasoning_content for kimi- ... es a concrete Kimi 400 after tool-call history. I did not run a live Kimi request in this read-only review. Automerge notes: - No ClawSweeper repair was needed after automerge opt-in. Validation: - ClawSweeper review passed for head 9a4605ee3843e2fdb9dd8d61b39400e58871923a. - Required merge gates passed before the squash merge. Prepared head SHA: 9a4605ee3843e2fdb9dd8d61b39400e58871923a Review: openclaw/openclaw#82550 (comment) Co-authored-by: Alex Knight <15041791+amknight@users.noreply.github.com>

Summary: - The PR preserves Kimi Coding reasoning_content replay for OpenAI-compatible tool-call follow-up turns, extends replay model-id matching, adds Kimi wrapper/tests, and updates the changelog. - Reproducibility: yes. at source level: current main drops or fails to synthesize reasoning_content for kimi- ... es a concrete Kimi 400 after tool-call history. I did not run a live Kimi request in this read-only review. Automerge notes: - No ClawSweeper repair was needed after automerge opt-in. Validation: - ClawSweeper review passed for head 9a4605e. - Required merge gates passed before the squash merge. Prepared head SHA: 9a4605e Review: openclaw#82550 (comment) Co-authored-by: Alex Knight <15041791+amknight@users.noreply.github.com>

openclaw-barnacle Bot added agents Agent runtime and tooling extensions: kimi-coding size: M maintainer Maintainer-authored PR labels May 16, 2026

amknight marked this pull request as ready for review May 16, 2026 11:29

clawsweeper Bot added the clawsweeper:automerge Maintainer opted this PR into bounded ClawSweeper-reviewed automerge label May 16, 2026

fix kimi coding tool replay

9a4605e

clawsweeper Bot force-pushed the ak/fix-kimi-coding-replay branch from fcc577e to 9a4605e Compare May 16, 2026 11:46

clawsweeper Bot merged commit f8b7008 into main May 16, 2026
113 checks passed

clawsweeper Bot deleted the ak/fix-kimi-coding-replay branch May 16, 2026 11:54

steipete mentioned this pull request May 16, 2026

Kimi model leaks internal monologue into visible assistant replies #82480

Closed

github-actions Bot mentioned this pull request May 16, 2026

📡 Upstream Digest — 2026-05-16 13:03 UTC curtismercier/openclaw-mods#871

Open

steipete mentioned this pull request May 16, 2026

fix(kimi): include reasoning_content on thinking tool replays #76073

Closed

This was referenced May 20, 2026

Subagent spawn fails: thinking enabled but reasoning_content missing #70563

Closed

Kimi K2.6 reasoning_content 400 regression in long conversations after LCM compaction (follow-up #70392) #71491

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix Kimi Coding tool-call replay#82550

Fix Kimi Coding tool-call replay#82550
clawsweeper[bot] merged 1 commit into
mainfrom
ak/fix-kimi-coding-replay

amknight commented May 16, 2026

Uh oh!

clawsweeper Bot commented May 16, 2026 •

edited

Loading

Uh oh!

amknight commented May 16, 2026

Uh oh!

clawsweeper Bot commented May 16, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

amknight commented May 16, 2026

Summary

Reproduction

Root Cause

Fix

Verification

Real behavior proof

Uh oh!

clawsweeper Bot commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

amknight commented May 16, 2026

Uh oh!

clawsweeper Bot commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

clawsweeper Bot commented May 16, 2026 •

edited

Loading

clawsweeper Bot commented May 16, 2026 •

edited

Loading